Effects of Advertising and Product Placement on ... - Semantic Scholar

2 downloads 254 Views 533KB Size Report
May 31, 2008 - boxes and DVRs allow continuous tracking of channel tuning, .... viewers' exposure to television program
Effects of Advertising and Product Placement on Television Audiences Kenneth C. Wilbur1, Michelle S. Goeree2, Geert Ridder3

May 31, 2008

Abstract: Digital video recorder proliferation and new commercial audience metrics are making television networks’ revenues more sensitive to audience losses from advertising. There is currently little understanding of how traditional advertising and product placement affect television audiences. In this paper, we estimate a random coefficients logit model of viewing demand for television programs, wherein time given to traditional advertising and product placements is the “price” of consuming a program. Our sample includes audience, advertising, and program characteristics from more than 10,000 network-hours of prime-time broadcast television from 2004 to 2007. We find that the median effect of a 10% rise in traditional advertising time is a 15% reduction in audience size. We find evidence to suggest that creative strategy and product category factors are important determinants of viewer response to traditional advertising. We find evidence that suggests product placement causes viewer switching. In sum, our results imply that networks should give price discounts to those advertisers whose ads are most likely to retain viewers’ interest throughout the commercial break. Keywords: Advertising, Advertisement Avoidance, Branded Entertainment, Choice Modeling, Endogeneity, Industrial Organization, Media, Product Placement, Television

1

Asssistant Professor of Marketing, USC Marshall School of Business, [email protected]. Asssistant Professor of Economics, University of Southern California, [email protected]. 3 Professor of Economics, University of Southern California, [email protected]. We thank the Financial Economics Institute at Claremont McKenna College for financial support and Kevin Hesla for excellent research assistance. All remaining errors are our own. 2

1 Television viewing is the dominant leisure activity in America. In a telephone survey Americans reported watching 2.6 hours of television per day, more than half of total leisure time.1 Other measures suggest time spent viewing is much higher. Nielsen Media Research estimates the average adult watches 4.9 hours of television per day.2 Television remains advertisers’ most important medium. In 2007 the television industry earned $67.8 billion in advertising revenues. Those revenues grew 35% from 2001 to 2007, and accounted for 48% of cumulative advertising expenditures. While some other advertising media (e.g., internet display advertising) grow at higher percentage rates due to smaller revenue bases, television grew more than any other medium between 2001 and 2007.3 Traditionally, broadcast television networks have provided viewers with nominally free programs in exchange for their attention, and sold that attention to advertisers based on program audience measurements. The structure of the industry suggests that viewers have a relative preference for programs or non-television activities over watching advertising. If this were not the case, networks would presumably refrain from producing such costly programming. The traditional television business model has been weakened by two recent trends. First, viewers are acquiring digital video recorders (DVRs), which enable them to easily fast-forward past advertisements in recorded and “near-live” programming. The DVR was introduced in 2000, and 23% of American households owned one in April 2008.2 Figure 1 shows that broadcast networks have responded to DVR growth in part by increasing product placements (“unskippable advertising”) in their shows by about 40% in the three years to March 2008. However, there is currently no research investigating how viewers respond to product placement. Second, improvements in audience tracking technologies have changed business practices. Digital cable boxes and DVRs allow continuous tracking of channel tuning, leading advertisers to demand increasingly granular data about how many viewers watched a particular ad, rather than the program during which the ad appeared. Since September 2007, ad deals have been based on programs’ average commercial minute rating,4 rather than program rating. Many analysts expect more granular advertisement ratings to be used in the future. 1

Source: Bureau of Labor Statistics, “American Time Use Survey,” 2006. Source: Data reported online at www.tvb.org. Accessed May 2008. 3 Source: TNS Media Intelligence custom report. In 2007 advertisers spent $28.0 billion on magazines, $26.2 billion on newspapers, $11.4 billion on internet display advertising, $3.9 billion on outdoor advertising, and $3.4 billion on radio. 4 A commercial minute is any minute (e.g., 8:12:00 p.m.-8:12:59 p.m.) in which a part of a commercial is aired. 2

2

Figure 1. Growth of Product Placements, 4/2005-5/2007 12

11.5

11

10.5

10

9.5 APR 2005 AUG 2005 DEC 2005 APR 2006 AUG 2006 DEC 2006 APR 2007 AUG 2007 DEC 2007 ln(Prime-Time Seconds of Product Placements on Major Broadcast Networks)

Thus viewers are more able to avoid advertisements than ever before. And networks are more likely to be financially penalized for advertisement avoidance than ever before. Our purpose in this paper, then, is to understand the effects of advertising and product placements on television audiences. This understanding is important in practice for several reasons. First, it can inform networks’ sales strategy, influencing which advertisers they seek to sell commercial time to. Second, it can influence networks’ pricing. It may be optimal to raise ad prices for advertisers whose ads cause larger audience losses than average, or offer discounts to advertisers whose ads cause smaller audience losses. Third, viewer welfare is directly enhanced if networks can reduce viewer disutility from advertising. And if this reduction raises networks’ advertising revenues, there may be an indirect effect on viewer welfare in the form of increased program investments. Our key findings are that a 10% increase in advertising time causes a median audience loss of about 15%. Audience reaction to individual advertisements seems to be driven by advertising content. In the next section, we discuss salient features of the industry and the recent academic literature.

1. Industry background and relevant Literature This paper is primarily related to two disparate strands of the literature: advertisement avoidance and television viewing demand.

3 Prior research documents strategies television viewers use to avoid commercials. Danaher (1995) investigated Nielsen Peoplemeter data in New Zealand and found that audience figures fell by a net 5% during ad breaks, due to a 10% audience loss to switching and a 5% audience gain from channel switchers leaving other channels. However, the context of the study was a three-channel environment in which simultaneous ad breaks were commonplace. Using Peoplemeter data from the Netherlands, Van Meurs (1998) found that channel switching decreased audience size during advertising breaks by a net 21.5%. Woltman Elpers, et al. (2003) used eye-tracking technology in the lab to find that subjects stopped watching 59.6% and 76.1% of all commercials. They found that commercial watching increases with entertainment content and decreases with information content. Tse and Chan (2001) called viewers at home immediately after commercial breaks ended and found that, of households watching television, 80.8% reported avoiding commercials or diverting their attention in some manner. These finding are buttressed by the large literature on advertising “wear-in” and “wear-out.” For example, Siddarth and Chattopadhyay (1997) found the probability that a household switches channels during a particular ad is “J-shaped” with a minimum at 14 exposures. One particularly interesting paper is Teixeira, Wedel, and Pieters (2008). These authors used eye-tracking data to estimate a model of a viewer’s decision to stop watching a commercial as a function of brand presence, prominence, duration, and location on the screen. They used the estimates to calculate optimal brand presence strategies, finding that pulsing the brand at the beginning and end of an ad minimized the probability of ad avoidance. Advertising avoidance notwithstanding, until September 2007 advertising sales contracts were based on program ratings. Program ratings were defined by Nielsen Media Research as the fraction of potential viewers watching a given program for at least five minutes within any fifteen-minute block. Thus, the forms of advertising avoidance that most directly impacted network revenues were switching channels or turning off the television, as these are the two strategies that can decrease a program rating. Quite separate from advertisement avoidance, there is a long literature on predicting viewer demand for television programs. Rust and Alpert (1984) were the first to use a discrete choice model to explain viewing behavior, demonstrating that contrary to previous findings, programs are important predictors of network audiences. More recently, Shachar and Emerson (2000) introduced cast demographic variables in viewing demand estimation and showed that

4 viewers are more likely to watch programs that feature people who are demographically similar to themselves. Crawford (2000) used variation across cable systems in two snapshots of data to identify the welfare effects to consumers of the 1992 Cable Act, finding that mandatory price reductions were offset by changes in bundle characteristics and service erosion. Goettler and Shachar (2001) estimated a multidimensional ideal point model to calibrate a model of optimal program scheduling, finding that networks’ adherence to scheduling “rules of thumb” (e.g. no situation comedies after 10 p.m.) was suboptimal. Anand and Shachar (2005) used data on viewers’ exposure to television program “tune-ins” and subsequent viewing choices to identify tune-in effectiveness. They found that tune-ins are informative in nature: they make viewers more likely to watch programs that confer high subjective utility, and more likely to avoid programs that confer low subjective utility. Yang, Narayan, and Assael (2006) estimate a model in which husbands and wives have joint latent viewing preferences, finding that wives’ viewing behavior depends more strongly on husbands’ viewing status than vice versa. A few studies have measured audience sensitivity to advertising levels found in various media, controlling for characteristics of media content. Kaiser and Wright (2006) estimated a two-sided equilibrium model of viewers and advertisers of women’s magazines, finding that ads increased reader utility of magazines. Wilbur (2008b) found that a highly-rated broadcast network lost about 25% of its audience in response to a 10% increase in advertising time. Depken and Wilson (2004) estimated magazine-specific audience responses to advertising and found substantial heterogeneity across magazines, including many positive and many negative effects. The process by which advertising leads to increased or decreased viewership/readership has not been fully explored, but could be a function of consumer demographics, media content, media usage, advertising content, and advertising intrusiveness. Goeree (2008) found that advertising exposure and impact varies across demographic groups and advertising media, so it seems reasonable to expect that advertising avoidance varies across consumer demographics and media. We also study audience responses to product placement. The first on-screen product placement occurred shortly after the invention of the movie, when in 1896 the Lumiere brothers filmed women washing clothes with Lever Brothers’ Sunlight Soap placed in a prominent position. Lever Brothers provided Swiss film distribution in exchange for the favorable treatment. A commonly cited successful placement was the appearance of Reese’s Pieces in the

5 film E.T. the Extraterrestrial, to which Hershey’s attributed a 65% rise in sales. Less commonly discussed is the placement of Coors Lite in the same film, to which no sales rise was attributed (Newell, Salmon and Chang 2006). Balasubramanian, Karrh, and Parwardhan (2006) review the behavioral literature on product placement, attributing the many discrepancies among published findings to brand, consumer, and placement heterogeneity, and the difficulty of recreating product placement stimuli in laboratory settings. An interesting framework is proposed by Russell (2002). She finds that placements have differential effects on consumers’ memory and brand attitudes. Obtrusive placements are most likely to be remembered, but they positively influence consumers’ attitude toward the brand only when they are congruent with the plot, and can harm brand attitudes when they are incongruent with the plot. These findings seemingly refute Ephron’s (2003) conjecture about product placement: “If you notice, it’s bad. But if you don’t, it’s worthless.” Finally, there is a large recent theoretical literature on two-sided media markets. Prominent among these papers is Anderson and Coate (2005), which shows that television markets can fail by providing too many ads when available programs are poor substitutes, or too few when advertisers’ profits are large relative to viewers’ disutility of ads. Dukes and Gal-Or (2003) model both the market for advertising sales and its subsequent effects on a product market. They show that media outlets can benefit by selling exclusive advertising, since this softens product-market competition and raises advertisers’ willingness to pay. Our paper is relevant to this literature insofar as our results inform the assumptions it makes about how viewers respond to advertising of various types. The literature is reviewed by Anderson and Gabscewicz (2006). To our knowledge, our study is the first to estimate the effect of product placements on viewer switching. In addition, we examine the responsiveness of television audiences to traditional advertising in a dataset that is an order of magnitude larger than any studied previously. We do not know of any prior work that estimates how audience responsiveness to advertising varies with advertisement characteristics. We conclude this section with a note about terminology. Traditionally, a rating is the fraction of all potential viewers who watched a given program. A share is the fraction of all viewers watching television who watched a given program. Our data measure program ratings, so we use this terminology throughout the paper to avoid confusion. Similarly, we use the term

6 “product placement” to refer to the inclusion of brands or products within television programs, which is sometimes called “branded entertainment,” “plugs,” or “tie-ins.” We refer to blocks of time sold to advertisers as “traditional advertising” or simply “advertising.” We use the terms “program” and “show” interchangeably. An “ad creative” is a set of visual and audio stimuli encoded in a video file.

2. A Model of Television Viewing Behavior In this section we describe our model of television viewing demand, and later estimate several variants of it. We follow previous literature by assuming that each television viewer watches one network at a time, and model program viewership in a discrete choice framework. Given the aggregate nature of our data, we use a random coefficients logit model similar to Berry, Levinsohn, and Pakes (1995, hereafter “BLP”). We index networks with n and programs with j. A viewer chooses from n 1...N t networks airing top-100 programs within half hour t. (Our audience datasource is a set of weekly “top 100” programs, described further in section 3.4.) Viewer utility is determined by time effects, program and network characteristics, and preference parameters. There exists a one-toone mapping from network-half hours (nt) to program-half hours (jt).5 The indirect utility viewer i derives from watching network n in half hour t is given by

uint

v( pnt , qnt ;

i

)

X nt

i

nt

(1)

int

where pnt is the number of seconds of product placements on network n during half-hour t, qnt is the number of seconds of traditional advertising on network n during half-hour t,

i

is a vector

of utility parameters, and v( pnt , qnt ; i ) is the marginal utility of advertising and product placement. We report results under several specifications for v( pnt , qnt ; i ) in section 5. The X nt vector contains program, network, and time data. We include in X nt program characteristics (genre, whether the airing was a new episode); network-day dummies, to capture networks’ historical strengths and weaknesses; half-hour effects, to allow television utility to vary over the course of the night; and season-week dummies, to allow the utility of watching television to vary over weeks and years. Many previous studies (e.g., Moshkin and Shachar 5

We could alternatively think of a consumer choosing a program-half hour combination. To be consistent with previous literature we model the viewer’s decision as choosing a network-half hour.

7 2004) demonstrate the importance of state dependence in television viewing, so we also include the network’s audience rating for the same day-half hour in each of the previous five weeks.6 In entertainment categories like television shows, observed product characteristics are often inadequate to describe product quality. For example, both Friends and Coupling (US) were half-hour situation comedies featuring 6-member casts of Caucasian actors, but Friends lasted eight seasons while Coupling (US) was canceled after 11 episodes. The

nt

term represents

characteristics of the program that are unobserved to the researcher but known to viewers, advertisers, and network producers. We include program dummies in the model to estimate unobserved program characteristics, and discuss

further in section 4.1.

nt

Equation 2 defines the distribution of the random utility parameters about their means.7 i i

,

i

(2)

~ N (0, I K )

i

i

represents viewer tastes that are not observed by the econometrician and is a K-dimensional

vector drawn from a multivariate standard normal distribution. The

term is a KxK matrix of

parameters to be estimated. The number of parameters K can be as large as the combined dimensions of

i

and

i

, but is typically chosen to be smaller, as estimation time increases

exponentially in K. We could include individual demographics drawn from population-level distributions in Equation (2), but given that we do not have meaningful variation in viewer demographics over markets or time, it is not clear that these effects would be separately identified from

. However, as we discuss in section 4, we could estimate this model separately

for different demographic groups and hence our parameter estimates could vary over demographic groups. If we impose the restriction

0 , we have specified a multinomial logit

model. The

int

term is a mean zero stochastic term distributed i.i.d. type I extreme value across

i, n, and t. Choices are invariant to multiplication by a viewer-specific constant, so we fix the standard deviation of 6

.

The rationale for including previous weeks’ audience ratings is that many programs are serial in nature, so previous weeks’ ratings are likely to predict demand for the current program. The nature of our data prevents us from using programs’ lead-in and lead-out audiences, as is the standard practice in controlling for viewing persistence, because we do not observe variation in audience rating over half-hours within a date/program. 7 Including random coefficients ensures that predicted switching patterns will be based on program characteristics, rather than based solely on similarity in program audience ratings.

8 We can rewrite the model as

uint where

nt

v( pnt , qnt ; )

X nt

nt

nt

int

(3)

int

captures the base utility every viewer derives from

network n at time t. The composite random shock,

int

int

, captures viewer preference

heterogeneity. Viewers may elect to watch a program outside the top 100, or engage in a non-television activity. The value of the best available alternative (the “outside option”) is given by

ui 0t

0t

(4)

i 0t

Given that we cannot identify relative utility levels, we normalize

0t

to zero.8 The conditional

probability that viewer i watches network n at time t is e

sint

nt

1

in t

e

nt

(5)

ilt

l

and the viewers for whom network n at time t maximizes utility are described by demographics in the set

Ant

{( i , (

int

)) | uint

n} .

uimt , m

(6)

The audience rating for network n at time t is then given by snt

(7)

sint dF ( ) An t

where F ( ) denotes the cumulative distribution function of

i

.

3. Data To estimate the model we use data from two sources: TNS Media Intelligence (TNS) and the Television Bureau of Advertising (TVB). The TNS data are extensive and contain program genre classifications, detailed traditional advertising data at the level of the individual commercial, and detailed product placements at the level of the individual product placement. The TVB data report television audience ratings at the date-network-program level for the top 100 national

8

We could fix

01

to zero and estimate the

functionally equivalent.

0t

for t

1 , but we instead include time effects in X nt , which is

9 programs that aired during “prime time” evening hours each week, during which networks earn 61% of their advertising revenues. Since programs typically change on half-hour increments, our unit of observation is the date-network-half-hour, e.g. January 1, 2007, ABC, 8:00-8:30 P.M. We discuss each component of the data in more detail, and present descriptive statistics in section 3.5.

3.1 Program Data Program characteristics data come from TNS and consist of program name, genre, network, and date of each airing. As we discuss in section 3.2, we observe each advertisement within each program, so we are able to construct approximate start and end times for each program-date. The networks in the dataset are ABC, CBS, CW, FOX, NBC, UPN, and WB. ABC, CBS, and NBC broadcast national programs 8-11 (all times are P.M., Eastern Standard Time), seven nights a week. FOX broadcasts national programs 8-10 on all seven nights. UPN broadcasted 8-10 Monday through Friday, and WB broadcasted 8-10 Sunday through Friday. WB and UPN merged and began broadcasting as the CW Network in September 2006. CW broadcast 8-10 Sunday through Friday in the 2006-07 season. FOX, WB, and UPN affiliates typically fill their 10-11 hour with local news or syndicated comedies (“re-runs”). FOX started a new network called My Network Television in 2006. However none of its program audiences were large enough to be included in the top 100 in any week of our sample. TNS assigns each program exclusively to one genre. Numerous studies (e.g. Rust and Alpert 1984, Goettler and Shachar 2001) illustrate the importance of program genre in predicting program viewing demand. Table 1 lists the genres ordered by the frequency of the network-halfhours in which they are programmed in the sample. Genres range from News Magazine to Wrestling. But the striking feature of the data is its relative lack of dispersion. Four genres— Drama/Adventure, Slice-of-Life, Situation Comedy, and Police/Suspense/Mystery—accounted for 76.4% of prime-time network program-hours. At the other end of the distribution, 30 genres account for just 7.02%.

Table 1. Genre Frequency

10 Genre Drama/Adventure Slice-of-Life (or "Reality") Situation Comedy Police/Suspense/Mystery Feature Film News Magazine Wrestling

Frequency 34.21% 16.08% 14.70% 11.45% 5.49% 5.11% 2.08%

Genre Game Show Professional Football - Game Award/Pageant/Parade/Celebration Variety - General Professional Baseball - Game College Football - Game Other

Frequency 1.66% 1.14% 1.04% 0.99% 0.98% 0.55% 4.51%

There were a few programs that appeared on more than one network over the course of the sample. For example, the Super Bowl appeared on FOX in 2005, ABC in 2006, and CBS in 2007. When this occurred, we defined a separate program-network for each instance of the program. While our unit of observation is a date-network-half-hour, a network occasionally aired more than one program per half-hour slot. This affected less than 1% of the half-hours in our sample and was usually related to sports programming. For example, a game ran longer than its scheduled timeslot, or a half-hour included both a “pre-game show” and part of a game (two separate programs for which we observe separate audience ratings). We therefore had to choose which program’s audience rating to assign to some date-network-half-hours shared by two programs. We followed a two-step procedure. If exactly one of the two programs did not appear in any other half-hours, then we assigned that program’s audience rating to the half-hour. If both programs spanned multiple half-hours, then we assumed the program that contained more advertising during the date-network-half-hour in question accurately reflected the true audience rating. It was never the case that neither program spanned multiple half-hours.

3.2. Traditional Advertising Data Advertising data are drawn from TNS Media Intelligence’s “Stradegy” database. This database provides advertisers, advertising agencies, and other marketers with “competitive advertising intelligence.” It is widely subscribed within industry. For all advertisements that aired during the sample period, we observe the brand advertised (e.g. Coca-Cola Classic), the network, start time, and length of the ad, and a name identifying the specific ad creative. In addition, TNS manually classified each brand as

11 belonging to a category (e.g. Regular Carbonated Soft Drinks), an industry (Beverages), a subsidiary (Coca-Cola USA) and a parent company (Coca-Cola Co.). Networks aired about 250,000 advertisements during our sample period. These included about 29,000 different ad creatives for 5,000 different brands spanning 350 categories in 50 industries. We construct daily network-half-hour measures of the frequency of the most common ad creatives and most commonly advertised product categories, and measure their effects on viewer utility. We report the results in section 5.2. We have data on the average price of a thirty second commercial for each program on each date. Networks report these date-program average advertising costs to TNS and Nielsen after their programs air, who then report the data to media buyers. These data allow media buyers to estimate costs of future media plans. If networks over-report these costs, they have a greater ability to give advertisers perceived discounts when negotiating ad prices. If they under-report these costs, they may limit their programs’ potential advertising demand. We are not aware of any evidence of systematic under- or over-reporting, perhaps because of the repeated nature of transactions in this industry. (These are not “rate card” data.) We do not observe advertisements networks aired for their upcoming programs (“tuneins” or “promos”), as TNS’ ad-recording software was not able to distinguish tune-ins from network programs. Time given to tune-ins is a potentially important omitted variable; we describe our endogeneity controls in section 4.1. A 2001 report found that networks aired 4:07 minutes of tune-ins per half-hour. This compared with 9:44 minutes of advertising at the time of the report, and tune-ins and traditional advertising time had a correlation of -0.31 (AAAA/ANA 2001).

3.3. Product Placement Advertising Data TNS Media Intelligence began recording product placement information on March 28, 2005. In their database, a product placement is a visual, audio, or audio-visual representation of a brand or product, whether explicit or implied. Common examples include detailed prize descriptions on a game show, a logo on the t-shirt of a reality show contestant, or a partially identifiable truck driven by a police officer in a dramatic series. For each product placement, we observe the brand placed, the brand characteristics defined in section 3.2, and the product placement characteristics listed in Table 2. In the median

12 placement, an identifiable product or package is shown in the foreground with no other brands or products on the screen. Products are integrated into the program in just 16% of all placements. As with advertising, we aggregate over placements to construct measures of product placement at the date-network-half-hour level.

Variable Verbal Only Direct Visual Only Implied Visual Only Verbal & Direct Visual Verbal & Implied Visual Appearance Product or Package shown Brand Name shown Brand Mark shown Billboard or Graphic Overlay No Visual Interaction Interaction w/ Real Life Persona Interaction w/ Fictional Character No Interaction Visual Brand Interaction Interaction Product Interaction (Proper Use) Type Product Interaction (Improper Use) No Interaction Integration Integration as a Prize or Reward

Notes

Type

Visibility

Clutter

Visual Location Length

17.0% 51.2% 24.5% 5.3% 2.0% 63.9% 11.4% 4.6% 3.1% 17.0% 21.6% 37.4% 41.0% 7.9% 33.2% 1.0% 57.8% 1.3%

Integrated Directly into Game/Contest Integrated Partially Into Game/Contest Integration as a Sponsorship Other Integration No Integration Fully Visible Partially Visible Not Applicable No Clutter Clutter

2.8% 1.3% 8.2% 2.6% 83.9% 39.8% 40.6% 19.6% 58.1% 24.9%

Not Applicable Foreground Background Not Applicable in Seconds

17.0% 60.7% 22.2% 17.0% 23.4 9.0 42.3 920.0

Mean Med. St.D. Max

Brand/product is clearly identifiable Brand/product is not clearly identifiable

E.g. a character wears a shirt with a Nike logo

Characters who successfully completed a game or contest were given the brand as a reward The brand/product was featured during the game/contest The brand/product was used during the game/contest The brand was presented as a sponsor of the program E.g. the brand was integrated with the plot of a dramatic program

At least 1 other brand/product appeared on screen during a visual product placement

Table 2. Product Placement Descriptive Statistics. Networks typically do not reveal placement terms, so no available datasource reports product placement prices. Our understanding of the industry is that product placements are sometimes paid in cash, sometimes bartered, and sometimes are not paid. Payment is more likely

13 when plot integration or character interaction occurs, in which case the integration or interaction almost always depicts the brand or product favorably. In the product placement data, we observe episode names for regular programs. Therefore, for the second and third television seasons in our data, we were able to construct an indicator of whether each episode had appeared previously in the television season. 9 We call this variable NewEps. It stands to reason that new program episodes are more attractive to viewers than previously-aired episodes (“re-runs”), so we use this information in predicting viewing demand.

3.4 Television Audience Data Only a handful of television viewing datasets have been available to academic researchers in the past 20 years. Most of those datasets contain individual viewers’ program choices over a limited number of days and programs. Our dataset contrasts with others in that we have an unusually large number of time periods and programs, but we do not have cross-sectional variation over individuals or markets. We collected our audience data from weekly “top 100” program lists found on the TVB website, www.tvb.org. Each list ordered the 100 highest-rated programs that week, and included the programs’ national audiences, as measured by Nielsen Media Research. These national audience estimates are the same measurements that underpinned sales transactions between networks and advertisers. Weekly top-100 program lists were available for three demographic groups in each of three 35-week television “seasons,” 2004-05, 2005-06, and 2006-07. Each season began on the third Monday of September and ended on the third Sunday in May. The demographic groups are those traditionally used to measure television audiences: adults 18-49, adults 25-54, and households. A unit of observation is a date-network-program, so we assign each date-program rating to the network-half-hours in which that program aired. We observe an audience rating for each demographic group whose top-100 list included that program.

9

We were not able to observe this for the first season since the product placement data sample did not begin until March 2005.

14 We do not observe a program’s audience rating if it falls short of the 100th-highest audience rating in the week it aired. This truncation issue affects 20% of the date-network-half hours in household audiences and 22% of observations in the other demographic groups.

3.5. Descriptive Statistics Table 3 displays advertising time, ad revenue, and audience descriptive statistics by network. CBS had the largest audience measured in households, by far, with an average rating of 8.24, followed by ABC (6.55) and NBC (6.33). Yet FOX led in advertisers’ most desired demographic, adults 18-49 (4.15), followed by CBS (3.93) and ABC (3.77). This lead in adults 18-49 yielded Fox’s premiere position in advertising revenues per half hour. WB and UPN had audiences and advertising revenues about half as large as the big four networks. It is surprising to compare the CW network’s performance to WB and UPN. CW had smaller audiences than either of its constituent networks, and lower average advertising revenues than UPN. It would appear that the WB-UPN merger was unsuccessful, unless it produced substantial unobserved program cost savings. ABC sold the most advertising time in the sample, with an average of 370 seconds of ads per half-hour. It was followed by WB (349), CW (340), UPN (334), NBC (318), CBS (309), and FOX (296). Networks’ standard deviations indicate that there was a great deal of dispersion around these means, with standard deviations about 25% as large as means of advertising time. Table 4 shows the raw correlations between the major variables. Notably, the correlation between advertising and product placement is close to zero. This may suggest product placements’ fit with program narrative is the primary determinant of how many product placements to include in a program, rather than revenue maximization. It is also notable that product placement time is positively correlated with audiences, with correlations ranging from 0.13 to 0.17. Slice-of-Life and Situation Comedy genres contain more advertising and product placement time than Drama/Adventure and Police/Suspense/Mystery programs. While audiences across demographic groups are highly correlated, genre preferences depend on demographics. Household-level audiences are more likely to watch police programs than adults 18-49 (0.18 correlation to 0.07), while adults 18-49 are more likely to watch reality programs than the households audience (0.14 to 0.05).

15

Network

Variable

# Observations

Mean

Std. Dev

Minimum

Maximum

ABC Ad seconds per half hour Ad revenues per half hour Household audience rating Adult 18-49 audience rating Adult 25-54 audience rating

4337 370 88 4337 1,767,606 1,604,670 4089 6.55 3.34 3976 3.77 2.29 4053 4.29 2.52

15 900 0 28,800,000 1.9 25.43 0.95 16.52 1.21 17.69

Ad seconds per half hour Ad revenues per half hour Household audience rating Adult 18-49 audience rating Adult 25-54 audience rating

4357 309 91 4357 1,787,250 1,713,632 3926 8.24 3.23 3896 3.93 2.12 3918 4.90 2.43

30 1890 0 38,200,000 2.3 42.6 1.05 35.2 1.28 37.1

CBS

CW Ad seconds per half hour Ad revenues per half hour Household audience rating Adult 18-49 audience rating Adult 25-54 audience rating

840 840 392 396 306

340 568,699 2.71 1.72 1.70

80 356,235 0.46 0.41 0.31

90 0 1.8 1.1 1.1

630 2,250,000 4.2 3 2.5

FOX Ad seconds per half hour Ad revenues per half hour Household audience rating Adult 18-49 audience rating Adult 25-54 audience rating

2763 296 80 2763 1,809,414 2,379,557 2484 6.20 4.38 2480 4.15 3.20 2479 4.38 3.52

40 1200 0 62,400,000 1.81 41.13 1.11 33.22 1.05 35.58

Ad seconds per half hour Ad revenues per half hour Household audience rating Adult 18-49 audience rating Adult 25-54 audience rating

4371 318 82 0.2013889 1305 4371 1,585,161 1,242,114 0 10,100,000 4176 6.33 2.35 1.91 16.4 4059 3.45 1.54 1.04 10 4126 4.03 1.73 1.3 11.4

Ad seconds per half hour Ad revenues per half hour Household audience rating Adult 18-49 audience rating Adult 25-54 audience rating

1398 1398 645 647 580

334 507,876 2.81 1.71 1.70

68 355,769 0.54 0.39 0.33

120 36000 1.81 1 1.02

720 1,922,400 4.71 3.22 3.22

Ad seconds per half hour Ad revenues per half hour Household audience rating Adult 18-49 audience rating Adult 25-54 audience rating

1679 1679 756 795 740

349 714,339 2.94 1.84 1.85

73 272,382 0.67 0.46 0.43

90 98,400 1.69 0.85 0.99

690 2,235,300 5.38 3 2.85

NBC

UPN

WB

Notes: an observation is a date-network-time slot combination. The advertising data are from TNS; the audience ranking data are from TVB. Advertising expenditures are in nominal dollars.

Table 3: Program and Advertising Descriptive Statistics

16 Ad Sec. PP Sec. Ad Doll. Ad Seconds 1.00 Product Placement Seconds -0.04 1.00 Ad Dollars 0.33 0.12 1.00 Household Rating 0.02 0.13 0.66 Adults 18-49 Rating 0.05 0.17 0.72 Adults 25-54 Rating 0.04 0.16 0.70 Drama/Adventure 0.02 -0.05 -0.03 Police/Suspense/Mystery -0.10 -0.08 0.02 Situation Comedy 0.06 0.04 -0.04 Slice-of-Life ("Reality") 0.03 0.30 0.08

HH A18-49 A25-54

1.00 0.92 0.96 0.02 0.18 -0.18 0.05

1.00 0.99 0.02 0.07 -0.11 0.14

1.00 0.04 0.11 -0.13 0.10

Table 4. Correlations between Key Variables

4. Estimation 4.1 Endogeneity The error term in the model is

nt

, which represents program characteristics that may be known

to the networks and viewers but are unobserved by the econometrician. We specify nt

where

j

j

(2)

nt

is the mean of unobserved characteristics for program j, and j denotes the program that

aired on network n at time t. (Remember that we have a one-to-one mapping from nt into jt, so we could equivalently write estimate

j

jt

in place of

nt

.) We use the serial nature of the data to

with program-specific fixed effects. The

nt

term represents deviations from this

mean over time periods in which the program airs. We include NewEps, network-weekday and season-week dummies in X nt to try to remove variation in

nt

due to episode quality,

networks’ historical schedule strength, and temporally variable factors like weather. After including program and network-time effects, the source of error in the model is nt

, which represents deviations from mean unobserved program utility. This structural error

term could capture unobserved temporal variation in program quality as some episodes of a program may be more entertaining than others. It could also capture variation in time given to tune-ins. We include lags of ad price per viewer in the viewer utility function as ad price per

17 viewer is likely to be correlated with tune-in seconds.10 It could also capture measurement error in the audience data. Television networks may know their programs’ and episodes’ quality, including those aspects captured in

nt

, and may take it into account when setting traditional advertising and

product placement levels. As a result we have a potential endogeneity problem in that advertising choices may be functions of

nt

. (However it should be noted that if networks had complete

information about programs’ and episodes’ quality, we likely would see a lower rate of new program failure in the data.) We use three sets of instruments to address remaining endogeneity issues: (1) lags of traditional advertising time, (2) lags of product placement time, and (3) functions of competitors’ program characteristics. Traditional advertising time and product placement seconds are autocorrelated (1-week correlations of 0.42 and 0.44, respectively), so lags are good proxies for current advertising time and product placements. Their exclusion from the viewer utility function is justified if networks are myopic when setting traditional advertising time and product placements. The intuition motivating the third set of instruments follows Goeree (2008) and is similar to that used by BLP to correct for endogeneity of price in differentiated products markets. Rivals’ program characteristics enter the network’s profit function and therefore influence the network’s choice of ad and product placement time, since the optimal amount of advertising to do on a program depends upon the characteristics of all of the programs aired by rivals. Hence, characteristics of rivals’ products and various combinations of these characteristics can be used to instrument for endogenous advertising in that they are correlated with advertising aired during program j but not with program j’s unobserved quality. These instruments are given by g jt , where g is the number of competing networks offering a program of characteristic g within date/half-hour t. The characteristics g that we consider are NewEps and genre effects. For the third set of instruments to be valid we must assume (as in BLP) that the demand unobservable, evaluated at the true parameter values, is mean independent of observed program characteristics.11

10

We originally treated lags of ad price per viewer as an instrument, but instrumental variables validity tests reported in Appendix 1 indicated that their exclusion from viewer utility was not justified. 11 Given that we observe variation in the programs offered in different time slots, these instruments are valid even though we estimate program fixed effects.

18 4.2. Identification We discuss informally what variation in the data identifies the parameters. Associated with each network-half hour is a mean utility,

nt

, which is chosen to match observed and predicted

audience ratings. Audience levels identify the show, network-day, week, and half-hour effects. Holding these characteristics constant, correlations between audience, advertising, and product placement over time identify the mean utility parameters associated with advertising and product placement. In practice we cannot estimate a separate dummy for every show in the sample, because some shows did not appear frequently enough to identify their own dummy. Thus we assign a show dummy to as many shows as possible, and the remaining shows are described by networkday, time, and genre effects. Some genre effects are dropped because they are highly collinear with the set of show dummies for the shows belonging to that genre. We are able to separately identify show effects from network-time effects because of the rich scheduling variation over the three-year sample period. Identification of the taste distribution parameters

relies on patterns of viewer

substitution between shows. While the means of show utility are identified by audience sizes, the standard deviations are identified by the “stickiness” of how those audience sizes change when faced with variation in show competition, advertising, and product placements on competing networks within the same half-hour.

4.3. Estimation The econometric technique follows recent studies of differentiated products, such as BLP (1995) and Nevo (2000). We estimate the model using the Generalized Method of Moments (GMM). The moments match the predicted demographic audience ratings for network-half hours nt to the corresponding observed ratings. To estimate the restricted model we follow Berry (1994). Setting

0 , the model in

section 2 reduces to a multinomial logit with an estimating equation of ln snt

ln s0t

nt

v( qnt , pnt ; )

X nt

j

nt

(9)

where s0 t is the audience rating of the outside good (or one minus the sum of the “inside” shares). To estimate the parameters we interact the error term

nt

with a set of instruments Z.

19 In our OLS results, Z includes program dummies, season-week dummies, network-weekday dummies, half-hour dummies, genre dummies, NewEps, five lags of audience share and ad price per viewer, product placement characteristics, and the observed data in v( pnt , qnt ; ) . In our instrumental variables (IV) specifications, we drop the observed data in v( pnt , qnt ; ) and add five lags of traditional advertising time and five lags of product placement time into Z. To estimate the full set of random coefficients, we add the competitors’ program characteristics described above to Z along with the lags of advertising and product placement time, and adopt the two-step estimator proposed by BLP. The first step is to match the model’s ( S tobs , ) that is the implicit solution to

predicted ratings to observed ratings. We solve for

S tobs st ( , ) 0 ,

(10)

where S tobs and s t are vectors of observed and predicted audience ratings respectively and represents the complete parameter set. For each guess of the parameter set 0 nt

initial guess

, calculate snt0 (

0 nt

, ) , construct a new guess r nt

these last two steps r times until

( S tobs , )

r 1 nt

1 nt

, we start with an

0 nt

(S, )

S ntobs , and repeat snt ( nt0 ) 14

( S tobs , ) is arbitrarily close to zero ( 10

our application). We then calculate the structural error term substituting

r nt

( S tobs , ) for

nt

in

. The

error term is given by nt

r nt

( )

( S tobs , ) ( v ( q nt , p nt ;

i

)

X nt

j

),

(11)

We search over parameter values to minimize the GMM objective function ˆ

where

{

nt

arg min( Z '

)'

} is the Nx1 error term, and

Z ' Z to get a consistent estimate of ˆ

(Z '

1

(Z '

),

(12)

is a weighting matrix. As an initial guess we set )( Z '

)' , which we use in the final

parameter estimation. The BLP estimation routine has the desirable property that it is linear in preference means, which greatly speeds computation by reducing the number of parameters that enter the objective function nonlinearly. However it is still nonlinear in the standard deviations of the preference distributions, and computation time increases exponentially with the number of nonlinear parameters to be estimated. We restrict the number of parameters interacting with

20 unobserved viewer heterogeneity to two: those multiplied by the terms pnt and qnt , so K=2. To simulate individual television viewers, we invert the Normal distribution at 500 multivariate Halton draws for each random utility parameter, and use antithetic acceleration to produce 500 more draws to reduce simulation variance.12 Thus our total number of simulated viewers is 1,000. The data we use in estimation is the final seven weeks of the 2004-05 season, since product placement data was not available until March 28, 2005; and weeks 6-35 of the 2005-06 and 2006-07 seasons, since we have five lags of audience and ad price per viewer in our utility specification. For the random coefficients logit results in section 5.3, we sampled 1000 candidate ˆ ’s in a grid search to find starting values, and to ensure we could not find a better solution than (12), as the GMM objective function is not globally concave. We found many local minima, but when drawn over the range of grid points we sampled, the objective function looks convex to the eye in both dimensions of

. Computation time was about five days on a 3.2 GHz computer

using serial processing.

5. Results In section 5.1 we estimate several versions of the multinomial logit (MNL) model using the method described in section 4.3. The ease of MNL estimation makes it helpful for specification testing. We first explore how the proposed endogeneity controls affect the results and explore how v( pnt , qnt ; ) should be characterized. In section 5.2, we explore some substantive questions, such as how viewer utility varies with advertisement category, ad creative, and program genre. Section 5.3 reports our random coefficient estimates and estimated audience elasticities of advertising.

5.1. Multinomial Logit Results In section 4.2 we proposed to use show dummies to control for unobserved program characteristics. Without show dummies, we would expect advertising responsiveness to be biased upward, since networks would include higher ad levels in programs with higher 12

For more on antithetic acceleration see Stern (1997, 2000). Geweke (1988) shows if antithetic acceleration is implemented during simulation, then the loss in precision is of order 1/N (where N are the number of observations), which requires no adjustment to the asymptotic covariance matrix.

21 unobserved quality. As table 5 shows, without show dummies we find both advertising and product placements have significant, positive effects on utility. When we add show dummies, the point estimates fall markedly, and advertising time is again significant, but this time with the opposite sign. Our other endogeneity control is instrumental variables for intertemporal variation in unobserved program characteristics and unobserved tune-in levels. We present the results of instrumental variables robustness checks in Appendix 1. To summarize, we found that lags of advertising and product placements are valid instruments for advertising and product placement time, but their use does not materially affect the estimates. We proceed with OLS estimation on efficiency grounds. Please see Appendix 1 for complete details.

Results without show dummies

Results with show dummies

Ad Seconds

1.15E-04 (3.00)

-9.77E-05 (3.03)

Product Placement Seconds

2.81E-05 (2.25)

-1.75E-05 (1.38)

0.73

0.82

Adjusted R

2

T-statistics in parentheses. Dependent variable is the log-transform of the audience rating among adults 18-49.

Table 5. Effects of program dummies on advertising responsiveness estimates.

Next we consider what function we should use for advertising and product placement utility. We know of no extant theory available to guide our selection. It seems reasonable to expect that marginal utility of advertising and product placement may be nonlinear. Wilbur (2008b) reports finding no evidence of nonlinearity in advertising utility, but our dataset is an order of magnitude larger. We followed two common procedures to select a functional form; both led to the same conclusion. First, we used splines with varying numbers of knots to estimate the shape of the advertising and product placement marginal utility functions. Second, we added powers of each term to a linear specification and stopped when the next power added was not statistically significant. Both methods indicated that v( pnt , qnt ; ) should be cubic in traditional

22 advertising time, and quadratic in product placement time. We found no evidence of interactions between advertising and product placement. Figure 2 shows the estimated marginal utility of advertising for each of the three demographic groups. We found that the households demographic group is less averse to advertising than the other two groups, but there is no apparent difference between adults 18-49 and adults 25-54. Ad utility is everywhere decreasing, with an inflection point at 407 seconds. Just 16% of observed ad levels exceed this inflection point. 0.000 0

100

200

300

400

500

600

-0.010 -0.020 -0.030 -0.040 -0.050 -0.060 -0.070 -0.080

-0.090 -0.100 Adults 18-49

Adults 25-54

Households

Figure 2. Estimated Marginal Utility of Ad Seconds

0.030

0.025

0.020

0.015

0.010

0.005

0.000 0

100

200

300 Adults 18-49

400

500

600

Adults 25-54

700

800

900

1000

Households

Figure 3. Estimated Marginal Utility of Product Placements

23 Figure 3 shows viewers’ estimated response to product placement seconds. Product placement utility is concave, but increasing and positive over most of the variable’s range. households appear to respond the most positively to product placement, followed by adults 25-54 and adults 18-49. While we have included several endogeneity controls, we wondered if the product placement results still may be biased. It could be the case that product placements are naturally accommodated by certain types of program scenes that contain unobserved characteristics that are attractive to viewers. For example, if high-budget program scenes are attractive to viewers, and contain increased levels of product placement, our finding of positive product placement utility could be spurious. Such content could vary over episodes in a program and therefore escape the control provided by our program dummies. This problem seems unlikely to affect our traditional advertising utility estimates since advertising content is not typically influenced by program content. We collected some additional data to investigate this possibility. We describe the procedure in detail in Appendix 2. The website TV.com aggregates viewers’ opinions about television program quality. We supplemented these data with information from our sample, so we were able to separately control for episode quality and product placement. When we did this, we found that the estimated effect of product placement time on utility was negative and significant. Variable Type Verbal Only Direct Visual Only Implied Visual Only Verbal & Direct Visual Verbal & Implied Visual Appearance Product or Package shown Brand Name shown Brand Mark shown Billboard or Graphic Overlay No Visual Interaction Interaction w/ Real Life Persona Interaction w/ Fictional Character No Interaction Visual Foreground Location Background

Point Est. (T-Stat) .06 (0.9) -.02 (0.7) -.03 (1.1) -.04 (1.0) -.04 (0.8) .02 (0.4) .05 (1.1) .05 (0.8) --.08 (1.3) -.04 (0.7) -.01 (0.6) --

Variable Integration Integration as a Prize or Reward Integrated Directly into Game/Contest Integrated Partially Into Game/Contest Integration as a Sponsorship Other Integration No Integration Visibility Fully Visible Partially Visible Not Applicable Clutter No Clutter Clutter Visual Brand Interaction Interaction Product Interaction (Proper Use) Type Product Interaction (Improper Use) No Interaction

Table 6. Product Placement Characteristics Estimates

Point Est. (T-Stat) .00 (0.1) .04 (1.0) -.09 (1.5) .02 (0.8) .04 (1.8) --.01 (0.5) -.02 (0.6) -.00 (0.0) -.06 (0.9) .04 (0.6) .08 (1.1) --

24 Adults 25-54 appear to have a nearly equal reaction to ads as adults 18-49. Households are the least valuable audience metric and seem to be least negatively affected by traditional advertising, and most positively affected by product placements. From here on, we focus on models estimated using audience data for adults 18-49, as they are the group traditionally valued most highly by advertisers. We include in X nt the product placement characteristics described in section 3.1. X nt includes x lnt , which represents the fraction of product placement seconds on network n during half-hour t that have characteristic l. In this way we are able to separately control for the amount of placements during the program and the types of placements observed. The estimates, reported in Table 6, do not indicate that product placement characteristics drive program viewing decisions.

5.2. Some Substantive Results In this section we report results from some multinomial logit regressions that exploit the richness of the data. The questions we consider are: how do the effects of advertising on audience sizes vary by product category? How do they vary by program genre? How do they vary by ad creative? In addition to the above questions, we investigated whether advertising utility varied over the course of the sample, but we did not find any evidence to show that it did. We similarly did not find evidence that advertising utility varies with advertisement length.

Ad Utility by Product Category To estimate the effect of product category advertising on utility we set

v( pnt , qnt ; )

pcnt

p c

qcnt

c

q c

(13)

c

where pcnt is the product placement time given to brands in category c on network n in half-hour t, and q cjt is the corresponding ad level. Table 7 shows the significant estimates of

p c

and

q c

. Our priors were that beer and

movie ads would be positively associated with utility, and possibly car ads. The results are consistent with these expectations, as the most liked category ads include movies, DVDs, light

25 beer, regular beer, diet soft drinks, and four automotive categories. More surprising was the appearance of finance-related categories, including banks, insurance, and financial services. We reviewed some of these ads to try to understand the results further. Our general sense is that these ads contain higher entertainment value and production budgets than the typical ad, perhaps because they must capture consumer attention for products that might not otherwise be enjoyable to think about. The most-liked categories were corporate computing and participatory sports, though both represent a very small share of total advertising dollars. Corporate computing was dominated by a highly entertaining branding campaign by IBM, while the highest-spending brand in participatory sports was 1-800-SKYDIVE. The estimates indicate that viewers are averse to advertising in a variety of categories. Many are low-involvement categories like toothpaste, candy, cookies, and mouthwash. Others may have negative product associations such as diapers, vegetable juices, bleach, or pharmacies. Prescription medications and wireless telecommunications are the two largest spending categories that negatively impact viewer utility. Prescription medication ad utility may be impacted by US Food and Drug Administration rules regarding disclosure of medication side effects.

Positive Category Advertising Effectsa Category Computers, Corporate Participatory Sports Light Beer & Ale Regular Beer & Ale Ice Cream Home Audio Equipment Financial Products and Services Cars, European Courier Services Cars, Domestic Motion Pictures Insurance Diet Carbonated Soft Drinks Pre-Recorded Video & DVDs Banks, S&Ls Light Trucks, Asian Light Trucks, Domestic a

Negative Category Advertising Effectsa

Point Est. % All Ad % All Ad (T-Stat) Seconds Dollars Category 0.0026 (3.0) 0.07% 0.11% Prescription Medications 0.0024 (2.8) 0.03% 0.06% Wireless Telecom Providers 0.0024 (2.8) 0.40% 0.78% Toothpaste & Whiteners 0.0023 (5.6) 0.13% 0.30% Stationery, Greeting Cards 0.0014 (2.2) 0.18% 0.23% Candy & Mints 0.0013 (2.2) 0.29% 0.39% Real Estate Agencies 0.0013 (2.6) 0.32% 0.36% Cookies & Crackers 0.0012 (2.7) 0.46% 0.62% Mouthwashes & Breath Fresheners 0.0012 (3.3) 0.35% 0.53% Diapers (Adult, Infant And Toddler) 0.0012 (2.7) 0.75% 0.98% Bleach & Fabric Softeners 0.0011 (4.5) 5.76% 6.28% Spectator Sporting Events 0.0010 (8.7) 0.31% 0.36% Shoe Stores 0.0009 (2.0) 0.50% 0.65% Apparel 0.0008 (2.0) 1.86% 1.92% Pharmacies 0.0007 (2.7) 0.70% 0.75% Vegetable Juices 0.0006 (2.0) 1.83% 2.04% 0.0005 (2.7) 2.48% 3.28%

Point Est. % All Ad % All Ad (T-Stat) Seconds Dollars -0.0002 (-2.7) 7.06% 6.33% -0.0003 (-2.6) 6.09% 6.05% -0.0007 (-2.1) 0.59% 0.57% -0.0009 (-3.4) 0.48% 0.37% -0.0011 (-2.9) 0.62% 0.53% -0.0012 (-3.0) 0.54% 0.46% -0.0012 (-2.3) 0.32% 0.28% -0.0013 (-2.9) 0.37% 0.30% -0.0016 (-2.3) 0.14% 0.11% -0.0018 (-2.4) 0.12% 0.10% -0.0019 (-2.5) 0.14% 0.09% -0.0019 (-2.0) 0.14% 0.11% -0.0020 (-2.1) 0.08% 0.09% -0.0025 (-3.0) 0.08% 0.07% -0.0027 (-2.3) 0.07% 0.06%

Only effects significant at the 95% confidence level are shown, for categories that spent >= .05% of total advertising dollars.

Table 7. Category Advertising Utility

26 Table 8 displays product placement category effects. They lend some credence to the possibility that our product placement results are affected by endogeneity. Three of the most liked product placements are estimated to be for pudding, beer, and soft drinks, which may correlate with depictions of social settings, while cosmetics may correlate with scenes containing female models. Positive Category Product Placement Effectsa Category Gelatins and Puddings Regular Beer & Ale Cosmetics & Beauty Aids Regular Carbonated Soft Drinks Sneakers Motion Pictures

a

Negative Category Product Placement Effectsa

Point Est. % All PP (T-Stat) Seconds Category 0.0014 (2.0) 0.10% Apparel 0.0007 (2.0) 0.42% Pre-Recorded Video 0.0006 (2.1) 0.46% Corporate Advertising 0.0003 (4.7) 10.37% Magazines 0.0002 (3.0) 3.14% Cars, Domestic 0.0002 (2.7) 1.53% Wireless Telecom Providers Internet Service Providers Credit Cards Prepared Dinners & Entrees Employment Agencies Medical Supplies

Point Est. % All PP (T-Stat) Seconds -0.0003 (-2.7) 1.64% -0.0003 (-4.6) 1.09% -0.0004 (-2.4) 0.53% -0.0005 (-2.0) 0.91% -0.0007 (-2.0) 0.48% -0.0007 (-4.2) 1.77% -0.0008 (-4.7) 0.89% -0.0012 (-2.4) 0.29% -0.0022 (-2.7) 0.09% -0.0054 (-3.3) 0.12% -0.0070 (-2.3) 0.04%

Only effects significant at the 95% confidence level are shown.

Table 8. Category Product Placement Utility

Ad Utility by Program Genre To estimate the effect of program genre on advertising utility we set v( pnt , qnt ; )

1ynt ( pnt

1y

pnt2

2y

qnt

3y

qnt2

4y

qnt3

5y

)

(14)

y

where 1 ynt indicates that network n aired a program of genre y during half-hour t. None of the genre effects were significant, with the exceptions of comedy’s effect on advertising utility and police/suspense/mystery’s effect on advertising and product placement utility, for which all of the

’s were significant at the 95% level. Figure 4 displays the estimated ad utilities during

comedy and police/suspense/mystery genre next to the general ad utility estimated in section 5.1. It appears that viewers respond more positively to ads in comedy programs, and very strongly negatively to ads in police/suspense/mystery programs. Figure 5 shows that estimated product placement sensitivity begins to fall and turn negative at high levels of product placement during police/suspense/mystery programs.

27

1.15 0.95 0.75 0.55 0.35 0.15 -0.05 0

50

100

150

200

250

300

350

400

450

-0.25 -0.45 Adults 18-49 Ad Utility (all ads) SitCom Genre * Ad Sec Utility Police/Suspense/Mystery Genre * AdSec Utility

Figure 4. Ad Utility by Genre

0.04 0.03 0.02 0.01

0.00 0

100

200

300

400

500

600

-0.01 -0.02 -0.03 Adults 18-49 PP Utility (All PP)

Police/Suspense/Mystery Genre * PP Utility

Figure 5. Product Placement Utility by Genre

Ad Utility by Advertising Creative Our final substantive question is how individual advertisements vary in their effect on viewer utility. We respecify ad utility as H

v( pnt , qnt ; )

pnt

1

pnt2

2

q0nt

3

q02nt

4

q03nt

qhnt

5

(15)

h

h 1

where qhnt is the number of ad seconds devoted to creative h on network n at time t,

h

is the

effect of creative h on utility, and q0nt is the amount of all ad time given to creatives that are not in the set 1…H. We choose ad creatives to include in H by following two steps. First, TNS creative names sometimes include an integer at the end, to indicate that the creative is a minor

28 departure from a previously-logged commercial for the same brand. Typically these departures are :15-second versions of a :30-second ad, or a change in on-screen text in an otherwise identical ad. We drop this integer to pool across variations within an ad creative, yielding about 24,000 ad creatives in the sample. Second, we define a dummy variable for each of the 350 ad creatives that occurred on television most frequently during our restricted sample period. Thus H=350. Each ad described by a creative-specific utility parameter appeared at least 42 times. Of the 350 ad creative parameters, 35 were estimated to be significant at the 95% confidence level. Table 9 displays the creative names, brands, parameter estimates, and tstatistics for each of those creatives. It also shows what fraction of all ads and ad dollars in the sample each creative accounted for. We interpret these results with some caution. We presume that most of the ad creatives in the sample have some effect on viewer utility, and we would be better able to detect those effects if we had individual-level viewing data. We are looking here at the tails of the distribution of ad creative utility. With those caveats in mind, it is interesting to note what these ads do not have in common. It does not appear that brand identity is a primary driver of ad creative utility, as two brands (Verizon Wireless and Old Navy) have ad creatives with significant positive effects, and ad creatives with significant negative effects. However, none of the estimated creative effects contradict the positive and negative category-specific effects presented earlier. We watched the ads in Table 9 to try to get a general sense of what creative elements drive the results. We noticed that ads with positive effect estimates tended to be upbeat and positive, and to feature actors that appeared younger than about 40 years old. One advertisement featured a popular celebrity (Denzel Washington) and another had a song from a popular band (Kings of Leon). Ad creatives with negative effect estimates were more likely to feature actors appearing older than forty, convey negative messages, and depict scenes of frustration. Some contained what could be subjectively termed annoying stimuli, such as intentionally bad dancing (“Push It/I365 Nextel Phone”), high-pitched, rapid speech (“1500 Whenever Mins/Cheerleader on Phone”) or actors using made-up words in conversation (“Men and Ellen at Reception are Gellin”).

29

a

Point Est.

% of

% ad

TNS Ad Creative Name

Brand

(T-Stat)

all ads

dollars

It's Ok To Look

Match.Com Dating Service

.005 (2.1)

.04%

.02%

Man Gets Locked Out In Bathrobe

Burlington Coat Factory Men

.005 (2.2)

.04%

.02%

Biggest Sale Of The Year

JC Penney

.004 (2.1)

.03%

.02%

No Title Assigned - #3726541

Boys & Girls Club/Psa

.003 (2.4)

.04%

.04%

Man Drives Family To Gaze At Stars

Toyota Trucks Sequoia

.003 (3.3)

.09%

.11%

Tunics/Women Dance On Boat

Old Navy Clothing Store

.003 (2.4)

.05%

.07%

Duck Helps Couple Get By

Aflac Medical Insurance

.003 (2.2)

.04%

.04%

Trainer Gets Pumped Up From Song

Verizon Wireless Service

.002 (2.3)

.05%

.05%

Vehicle Drives On Building Edges

Ford Trucks Edge

.002 (2.0)

.04%

.09%

Molly's Chambers/Couple Dances

Volkswagen Autos Jetta

.002 (3.0)

.06%

.10%

Truck Performs Seesaw Ramp Trick

Toyota Trucks Tundra

.002 (2.4)

.06%

.08%

Woman Wakes Up In The Dark

Lunesta Sleep Rx

-.001 (-2.2)

.08%

.15%

Woman Had Mysterious Symptoms

Requip Restless Legs Syndrm Rx

-.001 (-2.0)

.04%

.07%

People...To Do/Prescription Assistance

Humira Rheumatoid Arthritis Rx

-.001 (-2.6)

.04%

.07%

No Title Assigned - #3978969

Foundation/Better Lf/Psa

-.002 (-2.2)

.08%

.08%

Bubbles Flow Over Bottle & Teeth

Listerine Whitening Rinse

-.002 (-2.1)

.05%

.04%

Effortless Meticuless Fabuless

Target Disc Multi-Pdts

-.002 (-2.1)

.04%

.06%

Father Says He Got Hosed

Verizon Wireless Service

-.002 (-2.1)

.05%

.04%

Lust For Life/Women In Europe

Royal Caribbean Cruises

-.002 (-2.6)

.06%

.07%

Family Shareplan/Man Talks To Family

Verizon Wireless Service

-.002 (-2.3)

.05%

.04%

Women Walk Around City In Shorts

Old Navy Clothing Store

-.002 (-2.2)

.04%

.06%

Breast Meal/2Pc Meal/3 Strip Meal

KFC Restaurant

-.002 (-2.3)

.05%

.05%

Woman Acquires Boxes To Be Mailed

USPS.com

-.003 (-2.5)

.05%

.03%

Man Offers People Fast Relief

Zantac 150

-.003 (-2.1)

.03%

.03%

Man Works At Vineyard

Claritin Allergy Remedy

-.003 (-2.4)

.07%

.05%

National Sales Race

Nissan Autos Altima & Sentra

-.003 (-2.0)

.04%

.03%

Push It/I365 Nextel Phone

Sprint PCS Wireless Service

-.003 (-2.0)

.03%

.03%

Tuscan Garlic Chicken

Olive Garden Restaurant

-.003 (-2.1)

.06%

.05%

No Hassle Rewards/Man Skis In Summer

Capital One Mastercard & Visa

-.003 (-3.0)

.05%

.04%

No Hassle Rewards/D Spade Answers No

Capital One Mastercard & Visa

-.003 (-2.3)

.05%

.06%

People Rinse Their Mouth With Product

Listerine Mouthwash

-.003 (-2.2)

.04%

.03%

1500 Whenever Mins/Cheerleader On Phone

T-Mobile Wireless Service

-.003 (-4.3)

.09%

.09%

A Night In...Castle Giveaway/Letterbox

Disneyparks.Com Online

-.004 (-3.3)

.04%

.03%

The Difference Between Services

Blockbuster.Com Store Online

-.004 (-3.6)

.04%

.04%

Men & Ellen At Reception Are Gellin

Dr Scholls Massaging Gel Insoles

-.004 (-2.3)

.05%

.03%

a

Only ad creative effects significant at the 95% level are shown.

Table 9. Ad Creative Utility

30 The results suggest that, consistent with Woltman Elpers et al. (2003), creative strategy drives viewer acceptance of advertising. While we find these effects to be interesting, they are suggestive at best. We hope that future research will undertake a systematic content analysis of a large sample of advertisements, and correlate ad creative characteristics (like what appeals are used) with viewer propensity to switch during the ad (as in Teixeira, Wedel and Pieters 2008).

5.3. Random coefficients logit results In this section we present results from the full random coefficients logit model, including elasticities of advertising. Our advertising utility specification is given in equation (16). v( pnt , qnt ;

i

)

pnt

1i

pnt2

qnt

2

3i

qnt2

4

qnt3

We tried including random coefficients for all five elements of

i

5

(16)

, but this greatly increased the

objective function value and the function no longer appeared to be roughly convex. Given the high correlations among powers of pnt and qnt , it may be impractical to expect the data to separately identify more than one element of

i

for each variable.

The main results of the estimation are shown in Table 10. The model fit the data well, with a Pseudo R2 of 0.7499. The mean effects of advertising and product placement have the same signs as those estimated in the multinomial logit model, but are not estimated precisely. Most of the significant effects are those associated with the program dummies, networkday dummies, and half-hour dummies. Table 11 shows the highest estimated program effects. The top programs were American Idol, Desperate Housewives, Grey’s Anatomy, and Lost. Table 12 shows the network-weekday point estimates. The highest significant point estimate is NBC’s Thursday night, which is the only network-night we know of to be branded in recent years (“Must See TV”). The primary reason to estimate the random coefficients logit model is to produce audience elasticities of advertising that are free from the Independence of Irrelevant Alternatives problem. The advertising elasticities generated by this model are

snt qmt qmt snt

qnt snt qmt snt

vi sint (1 sint )dF ( i ), if m qnt vi sint simtdF ( i ), if m qnt

n

n .

(17)

31 Table X. Random Utility Parameter Estimates Regressor Ad Sec. (Ad Sec.)2 3 (Ad Sec.) PP Sec. (PP Sec.)2 NewEps 1-week lag s nt 2-week lag s nt 3-week lag s nt 4-week lag s nt 5-week lag s nt Constant

Coeff. Est. (T-Stat) St. Dev. Est. (T-Stat) -1.1E-3 (-0.2) 2.4E-6 (0.1) -1.0E-9 (-0.1) 3.0E-6 (0.0) -1.2E-8 (-0.6) 1.5E+0 (0.6) 4.3E-2 (0.5) 1.4E-2 (0.3) 9.9E-3 (0.1) 1.3E-2 (0.3) 1.4E-2 (1.2) -1.6E-1 (0.0)

0.1 (0.0)

0.3 (0.0)

GMM Objective Pseudo R2

8.4558 0.7499

Note: lags of s nt are the the previous week's audience rating on network n at the same weekday/half-hour as t

Table 10. Random Coefficients Logit Parameter Estimates Program

Coeff. Est. (T-Stat)

Program: American Idol Program: Desperate Housewives Program: Grey'S Anatomy Program: Lost Program: House Program: 20/20 Program: 24

6.8E-1 (2.2) 5.9E-1 (4.6) 5.3E-1 (5.4) 5.2E-1 (3.2) 4.4E-1 (1.7) 3.4E-1 (2.0) 3.3E-1 (2.6)

Table 11. Program Effect Estimates Regressor ABC-Mon ABC-Tue ABC-Wed ABC-Thu ABC-Fri ABC-Sat CBS-Sun CBS-Mon CBS-Tue CBS-Wed CBS-Thu CBS-Fri CBS-Sat FOX-Sun

Coeff. Est. (T-Stat) Regressor 0.18 (1.3) 0.13 (1.5) 0.11 (1.4) 0.15 (1.8) -0.23 (-1.8) -0.29 (-2.0) 0.25 (0.9) 0.48 (2.7) 0.20 (1.5) 0.15 (1.5) 0.45 (1.6) 0.04 (0.2) 0.08 (0.3) 0.49 (2.5)

Coeff. Est. (T-Stat)

FOX-Mon FOX-Tue FOX-Wed FOX-Thu FOX-Fri FOX-Sat NBC-Sun NBC-Mon NBC-Tue NBC-Wed NBC-Thu NBC-Fri NBC-Sat WB-Thu

0.24 (2.5) 0.20 (1.2) 0.27 (2.7) 0.11 (1.0) -0.31 (-1.9) 0.39 (0.7) 0.05 (0.3) 0.27 (2.6) 0.23 (0.7) 0.07 (0.8) 0.40 (3.1) -0.03 (-0.2) -0.13 (-0.7) -0.26 (-1.8)

Note: ABC-Sun was chosen to be the excluded night. With one exception (WB-Thu), all CW-, UPN-, and WB-Weekday interactions were dropped due to scarcity of top-100 audience observations on those nights.

Table 12. Network-Weekday Effects

32 Equation 17 contains sint , the probability that simulated individual i picks alternative n at time t, given a change in ad time qmt . Therefore substitution patterns are not driven by aggregate market shares irrespective of program characteristics, as in the multinomial logit model, but instead they are calculated as the aggregation of simulated discrete choices given the estimated distribution of preference heterogeneity. Also notable is that the model produces a different elasticity for each network-half hour. We focus below on advertising elasticities, as the results in Appendix 2 suggest that our product placement utility estimates may be biased. Table 13 displays the median estimated audience elasticities of advertising. In general, if a broadcast network increases its advertising time by 10%, its program audience will fall by about 15%. The cross-elasticities are roughly comparable in nature across the “inside” networks and the outside option, but since the market share of the outside option is much larger than the sum of the shares of the inside networks, this indicates that when viewers leave an audience in response to an additional advertisement, they usually turn away from broadcast television altogether (tuning to a cable network, for example) rather than switching to another top-100 program.

Network ABC CBS CW FOX NBC UPN WB

ABC -1.49 0.00 0.08 0.00 0.03 0.00 0.06 0.00 0.07 0.00 0.02 0.00 0.02 0.00

CBS 0.06 0.00 -1.53 0.00 0.03 0.00 0.06 0.00 0.06 0.00 0.02 0.00 0.02 0.00

CW 0.08 0.00 0.05 0.00 -1.49 0.00 0.00 0.04 0.00 -0.00 -0.00

FOX 0.06 0.00 0.05 0.00 0.03 0.00 -1.45 0.00 0.06 0.00 0.02 0.00 0.02 0.00

NBC UPN 0.06 0.06 0.00 0.00 0.05 0.05 0.00 0.00 0.03 -0.00 0.00 0.06 0.06 0.00 0.00 -1.51 0.06 0.00 0.00 0.02 -1.56 0.00 0.00 0.02 0.00 0.02 0.00

WB 0.09 0.00 0.03 0.00 -0.00 0.05 0.00 0.03 0.00 0.04 0.00 -1.51 0.00

Outside Option 0.04 0.00 0.05 0.00 0.02 0.00 0.05 0.00 0.04 0.00 0.02 0.00 0.03 0.00

a

Table entry i,j reports the estimated elasticity of option j's national audience in response to a 10% increase in network i's observed advertising level. Reported elasticities are the medians of the distribution of national audience elasticities over days and half-hours.

Table 13. Estimated Audience Elasticities of Advertising.

It is interesting to compare our elasticities estimates to those of Wilbur (2008b). He estimated a similar model using audimeter/diary audience data from multiple local markets. He found median own-elasticities of about -2.5 for the highly-rated networks in his sample, and larger elasticities for low-rated networks. Our elasticities are smaller and more homogeneous by

33 comparison. The difference in our estimates may perhaps be attributed to the unreliability of diary data, which places a much higher burden on the audience member than the Peoplemeter technology used to produce the audience data in this paper.

6. Discussion In light of the increasing importance of advertisement avoidance, we estimated a model of television viewing demand in which viewing decisions depend on program characteristics, scheduling factors, advertising time and characteristics, and product placement time and characteristics. Our key findings are that a 10% increase in advertising reduces a network’s audience by a median 15%, and audience responses to advertising seem to be driven by product category and advertising content. Our findings imply that networks ought to price discriminate among advertisers in order to maximize audience retention throughout their commercial breaks. There are three ways this could be done in practice. The simplest way would be to give ad price breaks to advertisers in categories which have traditionally been associated with high-utility ad creatives, such as beer and movies. Accordingly, higher prices could be charged to those advertisers in categories that historically cause higher audience losses. A more nuanced way to implement this would be to set up a system whereby advertisers have to submit their creatives to standardized tests of audience acceptance. For example, an ad creative could be vetted by an online consumer panel or inserted into network programming online (e.g., on Hulu.com). This testing could measure viewer response to the ad. Given enough consumers in the panel and a standard approach toward testing creatives, a formula could be devised to adjust the advertiser’s price. The attraction of this idea is that it would give advertisers an increased incentive to produce engaging advertising, and could possibly correct the currently unpriced externality in which one ad causes an audience loss that then harms later advertisers in the commercial break. The third approach would be to base ad prices on more granular television audience measurements, such as second-by-second ratings currently extractable from the universe of digital cable boxes and digital video recorders (Wilbur 2008a). This would give advertisers the strongest incentives to avoid causing audience losses, but is the most difficult to implement,

34 since ownership of the data reside with multiple parties with potentially conflicting interests, and the television industry has historically been slow to agree upon and implement common metrics. We view all three of these suggestions as realistic. The first can feasibly be implemented right away, while the second probably needs to be refined after a design and testing phase. The third suggestion is the most difficult to set up, but would have the most positive impact on the television industry’s collective health in the long run. It would also likely have the greatest effect on viewer welfare, which is consequential in an industry with such a large share of gross domestic leisure time. Like all models, ours has several limitations which suggest directions for future research. Primary among these is our seeming inability to estimate unbiased product placement effects on utility in the full sample. Another limitation is that we have not modeled consumer uncertainty about advertising and product placement time, as was done by Anand and Shachar (2004, 2005). This is difficult to do reliably using aggregate share data, but could possibly be done by applying the approaches of Chen and Yang (2007) or Musalem, Bradlow and Raju (forthcoming). Finally, while we have used the best audience data available to us, there is obviously scope for estimating a similar model using more granular data, such as commercial minute ratings or second-bysecond set-top box data.

35 Appendix 1. Instrument Validation. We follow Bound, Jaeger, and Baker (1995) and Staiger and Stock (1997) to validate the use of lags of ad seconds and product placement seconds as instruments for current ad seconds and product placement time. There are three steps to this procedure. The first step is to use F-tests to determine whether the proposed instruments jointly explain the endogenous variables. The second step is to use F-tests to determine whether the proposed instruments can be justifiably excluded from the viewer utility function. Third, we use a Hausman specification test to gauge the difference between the OLS and IV estimates. The first two steps of this procedure formalize the standard instrumental-variables intuition that valid instruments should be (1) correlated with the endogenous regressor, and (2) uncorrelated with the error term in the 2nd-stage equation. The third step checks whether IV estimation changes the point estimates of the endogenous regressors. If it does not change the estimates, we retain OLS estimates on efficiency grounds. Table A1 displays the results of the first two steps. The first column of the table presents results from the first-stage regression of advertising seconds on the proposed instruments and the exogenous variables in the viewer utility function. Lags of advertising seconds are significant and the F-test rejects the null hypothesis that the candidate instruments jointly do not explain the dependent variable at a high confidence level. The second column of the table indicates that the instruments are good predictors of Product Placement seconds. The third column tests the exclusion restrictions. The F-statistic fails to reject the null that the instruments do not jointly explain the log-transformed program ratings. Thus we conclude that lags of advertising and product placement time are valid instruments for current advertising and product placement time. The final step of the validation process is to compare the parameter estimates under OLS and IV. If there is no significant difference, OLS results are preferred on efficiency grounds. The Hausman test fails to reject the null that the OLS estimates are preferred to the IV estimates. Thus the use of these instruments does not change the estimated effects of the potentially endogenous variables enough to justify the loss of efficiency associated with IV estimation. We can think of three possible reasons why the data may reject the IV estimates. First, it could be that networks are unable to predict how their programs’ utility varies over time. This would be consistent with the high rate of new show failure observed in the industry. Second, it may be that networks know how their shows’ utility varies over time, but do not use this knowledge in setting advertising and product placements. This might be consistent with a mixed

36 strategy equilibrium. Third, it is possible that endogeneity affects our estimates, but we cannot correct for it without a stronger set of instruments. We explore this final possibility further in Appendix 2. We conclude that OLS estimation is appropriate in any of these three scenarios.

Instrument

First-stage Est. in AdSec Eqn. (T-Stat)

First-stage Est. in PP Eqn. (T-Stat)

2nd-Stage Est. in Viewer Demand Eqn. (T-Stat)

Ad Seconds 1st Lag 2nd Lag 3rd Lag 4th Lag 5th Lag Product Placement Seconds 1st Lag 2nd Lag 3rd Lag 4th Lag 5th Lag

0.16 (17.38) 0.13 (13.64) 0.09 (9.89) 0.09 (9.33) 0.09 (10.47)

-0.03 (-1.53) 0.01 (0.34) -0.02 (-0.99) 0.00 (0.2) 0.01 (0.62)

4.5E-5 (1.39) 4.4E-6 (0.14) 8.0E-6 (0.25) -1.7E-5 (-0.54) -3.3E-5 (-1.07)

2.0E-3 (0.6) 4.2E-3 (1.29) -5.7E-3 (-1.57) -2.5E-3 (-0.7) 4.1E-4 (0.11)

0.06 (7.18) 0.03 (3.5) 0.03 (3.28) 0.03 (3.81) 0.04 (4.53)

-2.9E-5 (-2.51) -2.5E-6 (-0.22) 1.7E-5 (1.4) -8.5E-6 (-0.68) 1.8E-6 (0.14)

Null Hypothesis

No joint effect

No joint effect

No joint effect

2

0.4383

0.6307

0.8315

2

0.3316 173.26 2.32 0 Reject Null

R in unrestricted model R in restricted model Joint Significance F-Stat 99% Critical Value P-Value Result

0.6231 0.8313 18.73 1.08 2.32 2.32 1.71E-34 0.37 Reject Null Don't Reject Null

Table A1. Instrumental Variables Validity Checks

37 Appendix 2. Estimating Product Placement Utility with Episode Quality Controls We are concerned that, despite our controls for endogeneity, the product placement utility estimates may be positively biased. It could be that occasions for product placement in a program episode correlate with unobserved episode characteristics that increase viewer utility. The program effects control for variation in unobserved program characteristics across programs in the sample, but not across episodes within a program. It is this latter variation that may be correlated with product placement, thereby creating the bias. If we had an independent measure of program quality that varied over episodes within a program, we could control for unobserved episode quality and measure the effect of product placement independently of episode quality. We were able to find such a measure in an online database, TV.com. This site states that “TV.com is home to millions of television fans contributing and connecting via their favorite shows. From program ratings and episode reviews to forum posts and blogs, the fans provide almost all of the site’s content…” The site is a wiki in which viewers can list, review, and assign quality “scores” to television programs and episodes. Each program and episode in the database has been scored on a 1-10 scale by TV.com users. Popular programs’ episodes are often rated by 400 or more users. The website says 72.3% of its users are between the ages of 18 and 49, 55% are female, and they watch an average of 21 television hours per week. Thus we think their ratings may be a reasonable proxy for episode utility experienced by adults 18-49. The benefit of using this dataset is that it contains average episode quality scores, which will allow us to control for the endogeneity issue discussed in the previous paragraph. The TV.com data are rich but incomplete and often inaccurate. The website’s program database is missing many of the programs in our sample and the quality score data is very sparse for lesser-known programs. The episode database seems to have even more problems. Even popular programs have duplicate episode listings and are missing some new episodes. Some episodes’ air dates are listed incorrectly. We also considered using data from tv.yahoo.com (another wiki), but this database seemed to have even more omissions and duplicates than TV.com. We therefore concluded it would be infeasible to add the episode-level TV.com data into the TNS/TVB sample. Instead, we decided to download some of the good TV.com data and supplement it with TNS/TVB data. The data seem to contain fewer flaws for popular programs, so we identified the most-frequently-programmed show in each of the four most common genres in the TNS/TVB

38 sample: 24, CSI, Scrubs, and Extreme Makeover. We then narrowed the airings of these programs to new episodes in the 2005-2006 and 2006-2007 seasons. We downloaded episode quality scores for all of the program/dates on which TV.com and TNS agreed a new episode appeared. This gave us a sample of 159 program-episodes. We supplemented this with episodelevel data from our TNS/TVB sample. We estimated a multinomial logit model at the program-episode level in which the dependent variable was the log-transformation of observed program/episode ratings and outside share data. The independent variables were the average number of advertising seconds per halfhour, average number of product placement seconds per half-hour, the average previous week’s audience in the same network-half hours to control for state dependence, a scalar indicating how many new episodes of the program had previously been aired in the same season to control for narrative arc, program dummies, a dummy for the 2005-06 season, and weekday dummies. We also tried including calendar-month dummies, product placement characteristics, powers of advertising and product placement time, and interactions between program dummies and product placement time, but none of these increased the adjusted R2 enough to justify the degrees of freedom they cost. Table A2 shows our results. The primary result of interest is the effect of product placement on utility, which is estimated to be negative and significant at the 95% confidence level. This finding is robust to basic specification changes. It suggests that our product placement estimates reported in section 5 are positively biased. The effects of advertising and episode quality score on utility are not estimated to be significant. Ad price per viewer, included to control for unobserved tune-in levels, is negative and significant. The weekday effects, season dummy, and program effects are all significant, yielding a high degree of fit. State dependence is not estimated to be significant, perhaps due to its correlation with the program and weekday dummies. In sum, we conclude that our product placement results in the full model are likely biased upwards. However, we are not able to control for this bias in the full sample.

39 Regressor Product Placement Seconds Ad Seconds TV.com Episode Score Ad Price per Viewer Previous week audience Episode Number 05-06 Season 24 Extreme Makeover Scrubs Monday Tuesday Thursay Constant 2

Adjusted R Number of observations

Coeff. Est. (T-Stat) -.0002 (-2.2) .0008 (1.3) -.0271 (-1.0) -1.8E-7 (-3.8) -.0005 (-0.1) -.0164 (-9.0) .1555 (5.3) -.6976 (-5.2) -1.1232 (-10.1) -.8832 (-17.3) -.2052 (-1.9) -.6860 (-6.8) -.5693 (-6.8) -1.5852 (-4.7) 0.8766 159

The dependent variable is ln(s nt )-ln(s 0t ), where s nt is audience share among Adults 18-49, and t indexes network-time periods in which valid episode rating data could be obtained from tv.com for the programs 24 , CSI , Extreme Makeover , and Scrubs .

Table A2. Program-Episode Utility Effects

40 References American Association of Advertising Agencies and Association of National Advertisers (AAAA/ANA). 2001. Television Commercial Monitoring Report. Mimeo. Anand, B., and R. Shachar. 2004. Brands as beacons: a new source of loyalty to multiproduct firms. Journal of Marketing Research, 41 (2): 135-150. Anand, B., and R. Shachar. 2005. Advertising, the Matchmaker. Mimeo, Tel Aviv University. Anderson, Simon P., Stephen Coate. 2005. Market Provision of Broadcasting: a Welfare Analysis. Review of Economic Studies, 72(4): 974-972. Anderson, Simon P., Jean Gabszewicz. 2006. The Media and Advertising: a Tale of Two-Sided Markets. Forthcoming in Handbook of Cultural Economics, eds. Victor Ginsburgh and David Throsby, Elsevier: North Holland. Balasubramanian, Siva K., James A. Karrh, Hemant Patwardhan. 2006. Audience Response to Product Placements. Journal of Advertising, 35(3): 115-141. Berry, Steven. 1994. Estimating Discrete Choice Models of Product Differentiation. Rand Journal of Economics, 25, 242-262. Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. Automobile Prices in Market Equilibrium. Econometrica, 63, 841-890. Bound, John, David A. Jaeger, and Regina M. Baker. 1995. Problems with Instrumental Variables Estimation when the Correlation between the Instruments and the Endogenous Explanatory Variable is Weak. Journal of the American Statistical Association, 90, 433450. Byzalov, Dmitri, Ron Shachar. 2004. The Risk Reduction Role of Advertising. Quantitative Marketing and Economics, 2: 283-320. Chamberlain, G. 1987. Asymptotic Efficiency in Estimation with Conditional Moment Restrictions. Journal of Econometrics, 34: 305-344. Chen, Yuxin and Sha Yang. 2007. Estimating disaggregate models using aggregate data via augmentation of individual choice. Journal of Marketing Research, 4: 596-613. Crawford, Gregory S. 2000. The Impact of the 2000 Cable Act on Household Demand and Welfare. RAND Journal of Economics, 31 (3): 422-449. Crawford, Gregory S. 2008. The Discriminatory Incentives to Bundle in the Cable Television Industry. Quantitative Marketing and Economics, 6(1): 41-78. Danaher, Peter J. 1995. What Happens to Television Ratings during Commercial Breaks? Journal of Advertising Research, 35(1): 37-42. Depken II, Craig A., Dennis P. Wilson. 2004. Is Advertising a Good or a Bad? Evidence from U.S. Magazine Subscriptions. Journal of Business, 77(2): 61-80. Dukes, Anthony, Esther Gal-Or. 2004. Negotiations and Exclusivity Contracts for Advertising. Marketing Science, 22 (2): 222-245. Ephron, Erwin. 2003. The Paradox of Product Placement. Mimeo. http://www.ephrononmedia. com/article_archive/article_pdf/placement_05_03.pdf Geweke, J. 1988. Antithetic Acceleration of Monte Carlo Integration in Bayesian Inference, Journal of Econometrics, 38: 73-89. Goeree, Michelle S. 2008. Limited Information and Advertising in the US Personal Computer Industry. Forthcoming Econometrica. Goettler, Ronald L., Ron Shachar. 2001. Spatial Competition in the Network Television Industry. RAND Journal of Economics, 32 (4): 624-656.

41 Goldberg, Marvin E., Gerald J. Gorn. 1987. Happy and Sad TV Programs: How They Affect Reactions to Commercials. Journal of Consumer Research, 14(3): 387-403. Kaiser, Ulrich and Julian Wright. 2006. Price Structure in Two-Sided Markets: Evidence from the Magazine Industry. International Journal of Industrial Organization, 24 (1): 1-28. Musalem, Andres, Eric T. Bradlow, Jagmohan S. Raju. 2007. Who’s got the coupon: estimating consumer preferences and coupon usage from aggregate information. Journal of Marketing Research, forthcoming. Nevo, A. (2000): A Practitioner's Guide to Estimation of Random Coefficients Logit Models of Demand, Journal of Economics and Management Strategy, 9, 513-548. Newell, Jay, Charles T. Salmon, Susan Chang. 2006. The Hidden History of Product Placement. Journal of Broadcasting & Electronic Media, 50(4): 575-594. Russell, Cristel Antonia. 2002. Investigating the Effects of Product Placements in Television Shows: the Role of Modality and Plot Connection Congruence on Brand Memory and Attitude. Journal of Consumer Research, 29: 306-318. Rust, Roland T., Mark I. Alpert. 1984. An Audience Flow Model of Television Viewing Choice. Marketing Science, 3 (2): 113-124. Shachar, Ron, John W. Emerson. 2000. Cast Demographics, Unobserved Segments, and Heterogeneous Switching Costs in a Television Viewing Choice Model. Marketing Science, 37 (2): 173-186. Shachar, Ron, Bharat Anand. 1999. The Effectiveness and Targeting of Television Advertising. Journal of Economics and Management Strategy, 7(3): 363-396. Siddarth, S., Amitava Chattopadhyay. 1998. To Zap or Not to Zap: a Study of the Determinants of Channel Switching during Commercials. Marketing Science, 17 (2): 124-138. Staiger, Douglas, James H. Stock. 1997. Instrumental Variables Regression with Weak Instruments. Econometrica, 65: 557-586. Stern, S. 1997. Simulation-Based Estimation. Journal of Economic Literature, 35, 2006-2039. Stern, S. 2000. Simulation Based Inference in Econometrics: Motivation and Methods, in Simulation-Based Inference in Econometrics: Methods and Applications, ed. by Mariano, Weeks, and Schuermann. Cambridge, U.K.: Cambridge University Press. Teixeira, Thales, Michel Wedel, Rik Pieters. 2008. Moment-to-Moment Optimal Branding in TV-Commercials: Preventing Avoidance by Pulsing. Marketing Science, forthcoming. Tse, Alan C. B., Chi-Fai Chan. 2001. Zapping Behavior during Commercial Breaks. Journal of Advertising Research, 41 (3): 25-30. Van Meurs, Lex. 1998. Zapp! A Study on Switching Behavior during Commercial Breaks. Journal of Advertising Research, 38 (1): 43-53. Wilbur, Kenneth C. 2008a. How the Digital Video Recorder Changes Traditional Television Advertising. Journal of Advertising, 38 (1): 143-149. Wilbur, Kenneth C. 2008b. A Two-Sided, Empirical Model of Television Advertising and Viewing Markets. Marketing Science, 27 (3). Woltman Elpers, Josephine L. C. M., Michel Wedel, and Rik G. M. Pieters. 2003. Why do Consumers Stop Viewing Television Commercials? Two Experiments on the Influence of Moment-to-Moment Entertainment and Information Value. Journal of Marketing Research, 40 (4): 437-453. Yang, Sha, Vishal Narayan, Henry Assael. 2006. Estimating the Interdependence of Television Program Viewership between Spouses: a Bayesian Simultaneous Equations Model. Marketing Science, 3, 336-349.